Upload model with sharded safetensors

.gitattributes

@@ -0,0 +1,5 @@
+model-00001-of-00005.safetensors filter=lfs diff=lfs merge=lfs -text
+model-00002-of-00005.safetensors filter=lfs diff=lfs merge=lfs -text
+model-00003-of-00005.safetensors filter=lfs diff=lfs merge=lfs -text
+model-00004-of-00005.safetensors filter=lfs diff=lfs merge=lfs -text
+model-00005-of-00005.safetensors filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,69 @@
+---
+license: apache-2.0
+language:
+- en
+library_name: transformers
+tags:
+- text-generation
+- causal-lm
+- transformer
+- argonne
+- pretrained
+pipeline_tag: text-generation
+---
+
+# Argonne 2.0
+
+A **4.9 billion parameter** decoder-only transformer language model trained from scratch.
+
+## Model Architecture
+
+| Component | Specification |
+|-----------|--------------|
+| **Parameters** | ~4.9B |
+| **Layers** | 24 transformer blocks |
+| **Hidden Size** | 4,080 |
+| **Attention Heads** | 24 query / 8 key-value (GQA) |
+| **Context Length** | 4,096 tokens |
+| **Vocabulary Size** | 151,665 |
+
+## Usage
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+
+model = AutoModelForCausalLM.from_pretrained(
+    "PursuitOfDataScience/Argonne-2.0",
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+    trust_remote_code=True
+)
+tokenizer = AutoTokenizer.from_pretrained("PursuitOfDataScience/Argonne-2.0", trust_remote_code=True)
+
+prompt = "The future of AI is"
+inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+outputs = model.generate(**inputs, max_length=256, do_sample=True, temperature=0.7)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+## License
+
+Apache 2.0
+
+## Citation
+
+```bibtex
+@misc{argonne2,
+  author = {PursuitOfDataScience},
+  title = {Argonne 2.0: A 4.9B Parameter Language Model},
+  year = {2026},
+  publisher = {Hugging Face},
+  url = {https://huggingface.co/PursuitOfDataScience/Argonne-2.0}
+}
+```
+
+## Links
+
+- GitHub: [PursuitOfDataScience](https://github.com/PursuitOfDataScience)
+- Hugging Face: [PursuitOfDataScience](https://huggingface.co/PursuitOfDataScience)

added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

config.json

@@ -0,0 +1,30 @@
+{
+  "architectures": [
+    "ArgonneModel"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "block_size": 4096,
+  "eos_token_id": 151645,
+  "hidden_dropout": 0.0,
+  "hidden_size": 4080,
+  "intermediate_size": 11008,
+  "max_position_embeddings": 4096,
+  "mlp_bias": false,
+  "model_type": "argonne2",
+  "n_embd": 4080,
+  "n_head": 24,
+  "n_layer": 24,
+  "num_attention_heads": 24,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 8,
+  "pad_token_id": 151643,
+  "rms_norm_eps": 1e-06,
+  "rope_theta": 10000.0,
+  "sliding_window": null,
+  "torch_dtype": "float32",
+  "transformers_version": "4.44.0",
+  "use_flash_attention": true,
+  "use_gradient_checkpointing": false,
+  "vocab_size": 151665
+}

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": 151645,
+  "pad_token_id": 151643,
+  "transformers_version": "4.44.0"
+}

model-00001-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d8d8c974aba12ebc3a6d89d854fe13145b6a12e0530bccc486071e3e4d15957
+size 5309647128

model-00002-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60ca8b990e3a42ee3d928d45e5a2c4c439b4e819c5743ed79cd1089e57bc62b6
+size 5209869352

model-00003-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de22769975a3f0760e6490c487f415fd877eacd3696deeccf5b697674cb44c59
+size 5209869376

model-00004-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e956c8268b484a8891958ed9ea663a92f50a4a06d8489492c8509d2df09017e5
+size 5351921856

model-00005-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fbb1caffb0be41117446f9ca814c4faae0688b9dcaf1146893f23b7ec163f1eb
+size 1066179000

model.py

@@ -0,0 +1,852 @@
+import math
+import importlib.util
+from bisect import bisect_left, bisect_right
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import (
+    AutoConfig,
+    AutoModel,
+    AutoModelForCausalLM,
+    PreTrainedModel,
+    PretrainedConfig,
+)
+from transformers.modeling_outputs import CausalLMOutput
+
+
+_flash_attn_available = importlib.util.find_spec("flash_attn") is not None
+if _flash_attn_available:
+    from flash_attn.flash_attn_interface import flash_attn_func
+
+
+class ArgonneConfig(PretrainedConfig):
+    """Configuration for the Argonne v2 family of models."""
+
+    model_type = "argonne2"
+
+    def __init__(
+        self,
+        vocab_size: int = 32000,
+        hidden_size: int = 4096,
+        num_hidden_layers: int = 48,
+        num_attention_heads: int = 32,
+        num_key_value_heads: Optional[int] = None,
+        intermediate_size: Optional[int] = None,
+        max_position_embeddings: int = 4096,
+        attention_dropout: float = 0.0,
+        hidden_dropout: float = 0.0,
+        rms_norm_eps: float = 1e-6,
+        rope_theta: float = 10000.0,
+        sliding_window: Optional[int] = None,
+        use_flash_attention: bool = True,
+        use_gradient_checkpointing: bool = False,
+        tie_word_embeddings: bool = True,
+        attention_bias: bool = False,
+        mlp_bias: bool = False,
+        pad_token_id: Optional[int] = None,
+        bos_token_id: Optional[int] = None,
+        eos_token_id: Optional[int] = None,
+        **kwargs,
+    ) -> None:
+        pad_token_id = pad_token_id if pad_token_id is not None else kwargs.pop("pad_token_id", None)
+        bos_token_id = bos_token_id if bos_token_id is not None else kwargs.pop("bos_token_id", None)
+        eos_token_id = eos_token_id if eos_token_id is not None else kwargs.pop("eos_token_id", None)
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs,
+        )
+        # Backwards compatibility with Argonne 1.x naming.
+        if "n_layer" in kwargs:
+            num_hidden_layers = kwargs["n_layer"]
+        if "n_head" in kwargs:
+            num_attention_heads = kwargs["n_head"]
+        if "n_embd" in kwargs:
+            hidden_size = kwargs["n_embd"]
+        if "block_size" in kwargs:
+            max_position_embeddings = kwargs["block_size"]
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = (
+            num_key_value_heads if num_key_value_heads is not None else num_attention_heads // 2
+        )
+        if self.num_key_value_heads < 1:
+            self.num_key_value_heads = 1
+        if num_attention_heads % self.num_key_value_heads != 0:
+            raise ValueError("num_attention_heads must be divisible by num_key_value_heads")
+
+        if intermediate_size is None:
+            width = int(8 * hidden_size / 3)
+            self.intermediate_size = ((width + 255) // 256) * 256
+        else:
+            self.intermediate_size = intermediate_size
+
+        self.max_position_embeddings = max_position_embeddings
+        self.attention_dropout = attention_dropout
+        self.hidden_dropout = hidden_dropout
+        self.rms_norm_eps = rms_norm_eps
+        self.rope_theta = rope_theta
+        self.sliding_window = sliding_window
+        self.use_flash_attention = use_flash_attention
+        self.use_gradient_checkpointing = use_gradient_checkpointing
+        self.tie_word_embeddings = tie_word_embeddings
+        self.attention_bias = attention_bias
+        self.mlp_bias = mlp_bias
+
+        if self.pad_token_id is None and self.eos_token_id is not None:
+            self.pad_token_id = self.eos_token_id
+
+        # Backwards compatibility aliases
+        self.n_embd = self.hidden_size
+        self.n_layer = self.num_hidden_layers
+        self.n_head = self.num_attention_heads
+        self.block_size = self.max_position_embeddings
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        orig_dtype = x.dtype
+        x = x.to(torch.float32)
+        # Clamp values to prevent overflow in pow(2)
+        x = torch.clamp(x, min=-65504.0, max=65504.0)
+        variance = x.pow(2).mean(-1, keepdim=True)
+        x = x * torch.rsqrt(variance + self.eps)
+        return (self.weight * x.to(orig_dtype))
+
+
+class RotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        max_position_embeddings: int = 2048,
+        base: float = 10000.0,
+        device: Optional[torch.device] = None,
+    ) -> None:
+        super().__init__()
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+
+        inv_freq = 1.0 / (
+            self.base
+            ** (torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim)
+        )
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._set_cos_sin_cache(max_position_embeddings, device or inv_freq.device, torch.get_default_dtype())
+
+    def _set_cos_sin_cache(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> None:
+        self.max_seq_len_cached = seq_len
+        t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype)
+        freqs = torch.outer(t, self.inv_freq)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
+        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
+
+    def forward(self, x: torch.Tensor, seq_len: int) -> Tuple[torch.Tensor, torch.Tensor]:
+        if seq_len > self.max_seq_len_cached:
+            self._set_cos_sin_cache(seq_len, x.device, x.dtype)
+        return (
+            self.cos_cached[:seq_len].to(dtype=x.dtype, device=x.device),
+            self.sin_cached[:seq_len].to(dtype=x.dtype, device=x.device),
+        )
+
+
+def rotate_half(x: torch.Tensor) -> torch.Tensor:
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    if position_ids is None:
+        cos = cos.unsqueeze(0).unsqueeze(0)
+        sin = sin.unsqueeze(0).unsqueeze(0)
+    else:
+        cos = cos[position_ids].unsqueeze(1)
+        sin = sin[position_ids].unsqueeze(1)
+
+    return (
+        (q * cos) + (rotate_half(q) * sin),
+        (k * cos) + (rotate_half(k) * sin),
+    )
+
+
+class GroupedQueryAttention(nn.Module):
+    def __init__(self, config: ArgonneConfig) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_kv_heads = config.num_key_value_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_groups = self.num_heads // self.num_kv_heads
+        self.sliding_window = config.sliding_window
+
+        self.q_proj = nn.Linear(
+            self.hidden_size,
+            self.num_heads * self.head_dim,
+            bias=config.attention_bias,
+        )
+        self.k_proj = nn.Linear(
+            self.hidden_size,
+            self.num_kv_heads * self.head_dim,
+            bias=config.attention_bias,
+        )
+        self.v_proj = nn.Linear(
+            self.hidden_size,
+            self.num_kv_heads * self.head_dim,
+            bias=config.attention_bias,
+        )
+        self.o_proj = nn.Linear(
+            self.num_heads * self.head_dim,
+            self.hidden_size,
+            bias=config.attention_bias,
+        )
+        self.o_proj._is_residual = True
+
+        self.attention_dropout = config.attention_dropout
+        self.use_flash_attention = config.use_flash_attention
+
+    def _repeat_kv(self, x: torch.Tensor) -> torch.Tensor:
+        if self.num_key_value_groups == 1:
+            return x
+        bsz, num_kv, seqlen, head_dim = x.shape
+        x = x[:, :, None, :, :].expand(bsz, num_kv, self.num_key_value_groups, seqlen, head_dim)
+        return x.reshape(bsz, num_kv * self.num_key_value_groups, seqlen, head_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        bsz, seqlen, _ = hidden_states.shape
+
+        query = self.q_proj(hidden_states)
+        key = self.k_proj(hidden_states)
+        value = self.v_proj(hidden_states)
+
+        query = query.view(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2)
+        key = key.view(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        value = value.view(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query, key = apply_rotary_pos_emb(query, key, cos, sin)
+
+        key = self._repeat_kv(key)
+        value = self._repeat_kv(value)
+
+        use_flash_attn_2 = (
+            _flash_attn_available
+            and self.use_flash_attention
+            and attention_mask is None
+            and query.dtype in (torch.float16, torch.bfloat16)
+            and self.head_dim % 4 == 0
+        )
+        use_scaled_dot = (
+            hasattr(F, "scaled_dot_product_attention")
+            and self.use_flash_attention
+            and query.dtype in (torch.float16, torch.bfloat16)
+            and self.head_dim % 4 == 0
+        )
+
+        attn_output = None
+        if use_flash_attn_2:
+            try:
+                flash_dropout = self.attention_dropout if self.training else 0.0
+                window = (
+                    (self.sliding_window, self.sliding_window)
+                    if self.sliding_window is not None
+                    else (-1, -1)
+                )
+                q = query.transpose(1, 2).contiguous()
+                k = key.transpose(1, 2).contiguous()
+                v = value.transpose(1, 2).contiguous()
+                attn_output = flash_attn_func(
+                    q,
+                    k,
+                    v,
+                    dropout_p=flash_dropout,
+                    softmax_scale=None,
+                    causal=True,
+                    window_size=window,
+                ).transpose(1, 2)
+            except RuntimeError:
+                attn_output = None
+
+        if attn_output is None and use_scaled_dot:
+            try:
+                # Use is_causal=True when no attention_mask (faster Flash Attention path)
+                # When attention_mask is provided, we need to combine it with causal masking
+                if attention_mask is None:
+                    attn_output = F.scaled_dot_product_attention(
+                        query,
+                        key,
+                        value,
+                        attn_mask=None,
+                        dropout_p=self.attention_dropout if self.training else 0.0,
+                        is_causal=True,
+                    )
+                else:
+                    # With attention_mask: need to pass it explicitly (slower but correct)
+                    # attention_mask should be 4D: (bsz, 1, seq, seq) or broadcastable
+                    attn_output = F.scaled_dot_product_attention(
+                        query,
+                        key,
+                        value,
+                        attn_mask=attention_mask,
+                        dropout_p=self.attention_dropout if self.training else 0.0,
+                        is_causal=False,  # Mask already includes causal component
+                    )
+            except RuntimeError:
+                # Fallback to math attention when kernels are unavailable
+                attn_output = None
+
+        if attn_output is None:
+            scores = torch.matmul(query, key.transpose(2, 3)) / math.sqrt(self.head_dim)
+            # Apply causal mask - use large negative instead of -inf for numerical stability
+            causal_mask = torch.triu(
+                torch.ones(seqlen, seqlen, dtype=torch.bool, device=hidden_states.device),
+                diagonal=1,
+            )
+            mask_value = -65504.0  # Large negative instead of -inf
+            scores = scores.masked_fill(causal_mask, mask_value)
+            # Apply attention_mask if provided
+            if attention_mask is not None:
+                scores = scores + attention_mask
+            attn_weights = torch.softmax(scores, dim=-1, dtype=torch.float32).to(query.dtype)
+            attn_weights = F.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+            attn_output = torch.matmul(attn_weights, value)
+
+        attn_output = (
+            attn_output.transpose(1, 2)
+            .contiguous()
+            .view(bsz, seqlen, self.num_heads * self.head_dim)
+        )
+        return self.o_proj(attn_output)
+
+
+class SwiGLUMLP(nn.Module):
+    def __init__(self, config: ArgonneConfig) -> None:
+        super().__init__()
+        self.gate_proj = nn.Linear(
+            config.hidden_size,
+            config.intermediate_size,
+            bias=config.mlp_bias,
+        )
+        self.up_proj = nn.Linear(
+            config.hidden_size,
+            config.intermediate_size,
+            bias=config.mlp_bias,
+        )
+        self.down_proj = nn.Linear(
+            config.intermediate_size,
+            config.hidden_size,
+            bias=config.mlp_bias,
+        )
+        self.down_proj._is_residual = True
+        self.dropout = nn.Dropout(config.hidden_dropout)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # Clamp intermediate values to prevent overflow
+        gate = self.gate_proj(x)
+        gate = torch.clamp(gate, min=-65504.0, max=65504.0)
+        up = self.up_proj(x)
+        up = torch.clamp(up, min=-65504.0, max=65504.0)
+        return self.dropout(self.down_proj(F.silu(gate) * up))
+
+
+class Block(nn.Module):
+    """Transformer block with GQA attention and SwiGLU feed-forward."""
+
+    def __init__(self, config: ArgonneConfig, layer_idx: int = 0) -> None:
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.attn = GroupedQueryAttention(config)
+        self.input_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.mlp = SwiGLUMLP(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.input_norm(hidden_states)
+        hidden_states = self.attn(hidden_states, position_embeddings, attention_mask)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.post_norm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class ArgonneModel(PreTrainedModel):
+    config_class = ArgonneConfig
+    _no_split_modules = ["Block"]
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config: ArgonneConfig) -> None:
+        super().__init__(config)
+        self.config = config
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.blocks = nn.ModuleList([Block(config, idx) for idx in range(config.num_hidden_layers)])
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = RotaryEmbedding(
+            config.hidden_size // config.num_attention_heads,
+            max_position_embeddings=config.max_position_embeddings,
+            base=config.rope_theta,
+        )
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        if config.tie_word_embeddings:
+            self.lm_head.weight = self.embed_tokens.weight
+
+        self.gradient_checkpointing = config.use_gradient_checkpointing
+        self.pipeline_partitions: Optional[List[Tuple[int, int, torch.device]]] = None
+        self.devices: List[torch.device] = []
+        self.output_device: torch.device = self.embed_tokens.weight.device
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.embed_tokens
+
+    def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
+        self.embed_tokens = new_embeddings
+        self.config.vocab_size = new_embeddings.num_embeddings
+        if self.config.tie_word_embeddings:
+            self.lm_head.weight = self.embed_tokens.weight
+
+    def get_output_embeddings(self) -> nn.Module:
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings: nn.Module) -> None:
+        self.lm_head = new_embeddings
+        if isinstance(new_embeddings, nn.Linear):
+            self.config.vocab_size = new_embeddings.out_features
+
+    def tie_weights(self) -> None:
+        if self.config.tie_word_embeddings:
+            self.lm_head.weight = self.embed_tokens.weight
+
+    def _init_weights(self, module: nn.Module) -> None:
+        if isinstance(module, nn.Linear):
+            std = self.config.hidden_size ** -0.5
+            if hasattr(module, "_is_residual"):
+                std = (2 * self.config.num_hidden_layers) ** -0.5
+            nn.init.normal_(module.weight, mean=0.0, std=std)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.hidden_size ** -0.5)
+
+    def set_gradient_checkpointing(self, enabled: bool = True) -> None:
+        self.gradient_checkpointing = enabled
+
+    def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None) -> None:
+        self.set_gradient_checkpointing(True)
+
+    def gradient_checkpointing_disable(self) -> None:
+        self.set_gradient_checkpointing(False)
+
+    def distribute_model(self, device_ids: Optional[List[str]] = None) -> None:
+        if device_ids is None:
+            num_gpus = torch.cuda.device_count()
+            if num_gpus < 1:
+                raise ValueError("No CUDA devices available for distribution.")
+            device_ids = [f"cuda:{i}" for i in range(num_gpus)]
+
+        if not device_ids:
+            raise ValueError("device_ids must contain at least one device identifier.")
+
+        self.devices = [torch.device(d) for d in device_ids]
+        num_blocks = len(self.blocks)
+
+        if num_blocks == 0:
+            raise ValueError("The model has no transformer blocks to distribute.")
+
+        block_param_bytes: List[int] = []
+        for block in self.blocks:
+            size_bytes = 0
+            for param in block.parameters():
+                size_bytes += param.numel() * param.element_size()
+            block_param_bytes.append(size_bytes)
+
+        block_cumsum: List[int] = [0]
+        for size in block_param_bytes:
+            block_cumsum.append(block_cumsum[-1] + size)
+
+        embed_bytes = sum(p.numel() * p.element_size() for p in self.embed_tokens.parameters())
+        rotary_bytes = sum(p.numel() * p.element_size() for p in self.rotary_emb.parameters())
+        norm_bytes = sum(p.numel() * p.element_size() for p in self.norm.parameters())
+        head_dtype_size = self.embed_tokens.weight.element_size()
+        head_bytes = self.config.hidden_size * self.config.vocab_size * head_dtype_size
+        if self.config.tie_word_embeddings and len(self.devices) == 1:
+            head_bytes = 0
+
+        total_bytes = (
+            block_cumsum[-1]
+            + norm_bytes
+            + head_bytes
+            + embed_bytes
+            + rotary_bytes
+        )
+        per_device_target = total_bytes / len(self.devices)
+
+        per_device_counts: List[int] = [0] * len(self.devices)
+        prev_cut = 0
+        for idx, _ in enumerate(self.devices):
+            remaining_devices = len(self.devices) - idx
+            remaining_blocks = num_blocks - prev_cut
+            if remaining_blocks <= 0:
+                per_device_counts[idx] = 0
+                continue
+            if remaining_devices == 1:
+                cut = num_blocks
+            else:
+                reserve = max(0, min(remaining_devices - 1, remaining_blocks - 1))
+                max_cut = prev_cut + (remaining_blocks - reserve)
+                lo = prev_cut + 1
+                device_overhead = 0
+                if idx == 0:
+                    device_overhead = embed_bytes + rotary_bytes
+                available_block_bytes = per_device_target - device_overhead
+                if available_block_bytes <= 0:
+                    cut = lo
+                else:
+                    target_total = block_cumsum[prev_cut] + available_block_bytes
+                    cut = bisect_right(block_cumsum, target_total, lo=lo, hi=max_cut + 1) - 1
+                    if cut < lo:
+                        cut = lo
+            per_device_counts[idx] = cut - prev_cut
+            prev_cut = cut
+
+        def compute_device_block_bytes() -> List[int]:
+            device_block_bytes: List[int] = []
+            cursor = 0
+            first_partition_idx = next(
+                (i for i, count in enumerate(per_device_counts) if count > 0),
+                0,
+            )
+            for idx, block_count in enumerate(per_device_counts):
+                if block_count <= 0:
+                    device_block_bytes.append(0)
+                    continue
+                next_cursor = min(cursor + block_count, num_blocks)
+                block_bytes = block_cumsum[next_cursor] - block_cumsum[cursor]
+                if idx == first_partition_idx:
+                    block_bytes += embed_bytes + rotary_bytes
+                device_block_bytes.append(block_bytes)
+                cursor = next_cursor
+            if len(device_block_bytes) < len(self.devices):
+                device_block_bytes.extend(
+                    [0] * (len(self.devices) - len(device_block_bytes))
+                )
+            return device_block_bytes
+
+        output_payload = norm_bytes + head_bytes
+
+        device_block_bytes = compute_device_block_bytes()
+        positive_indices = [i for i, count in enumerate(per_device_counts) if count > 0]
+        if positive_indices:
+            last_idx = positive_indices[-1]
+            while True:
+                if per_device_counts[last_idx] <= 1:
+                    break
+                other_indices = positive_indices[:-1]
+                if not other_indices:
+                    break
+                other_loads = [device_block_bytes[i] for i in other_indices]
+                max_other = max(other_loads) if other_loads else 0
+                if max_other == 0:
+                    break
+                last_load_with_head = device_block_bytes[last_idx] + output_payload
+                if last_load_with_head <= max_other:
+                    break
+                prev_idx = other_indices[-1]
+                if per_device_counts[prev_idx] <= 0:
+                    break
+                per_device_counts[last_idx] -= 1
+                per_device_counts[prev_idx] += 1
+                device_block_bytes = compute_device_block_bytes()
+                positive_indices = [
+                    i for i, count in enumerate(per_device_counts) if count > 0
+                ]
+                last_idx = positive_indices[-1]
+
+        device_block_bytes = compute_device_block_bytes()
+        positive_indices = [i for i, count in enumerate(per_device_counts) if count > 0]
+        last_active_idx = positive_indices[-1] if positive_indices else 0
+
+        partitions: List[Tuple[int, int, torch.device]] = []
+        start_idx = 0
+        for device, block_count in zip(self.devices, per_device_counts):
+            if block_count <= 0 or start_idx >= num_blocks:
+                continue
+            end_idx = min(start_idx + block_count, num_blocks)
+            for block in self.blocks[start_idx:end_idx]:
+                block.to(device)
+            partitions.append((start_idx, end_idx, device))
+            start_idx = end_idx
+
+        if not partitions:
+            partitions.append((0, num_blocks, self.devices[0]))
+            if per_device_counts:
+                per_device_counts[0] = num_blocks
+                if not device_block_bytes:
+                    device_block_bytes.append(block_cumsum[num_blocks])
+            if not device_block_bytes:
+                device_block_bytes = [block_cumsum[num_blocks]]
+
+        self.pipeline_partitions = partitions
+        self.output_device = partitions[-1][2]
+        output_device_idx = last_active_idx
+
+        first_device = partitions[0][2]
+        self.embed_tokens = self.embed_tokens.to(first_device)
+        self.rotary_emb = self.rotary_emb.to(first_device)
+        self.norm = self.norm.to(self.output_device)
+
+        if self.config.tie_word_embeddings and len(self.devices) > 1:
+            untied_head = nn.Linear(self.config.hidden_size, self.config.vocab_size, bias=False)
+            untied_head.to(self.output_device)
+            with torch.no_grad():
+                untied_head.weight.copy_(self.embed_tokens.weight.to(self.output_device))
+            self.lm_head = untied_head
+            self.config.tie_word_embeddings = False
+        else:
+            self.lm_head = self.lm_head.to(self.output_device)
+
+        print(f"Model distributed across {len(self.devices)} devices.")
+        running = 0
+        for idx, (block_count, device) in enumerate(zip(per_device_counts, self.devices)):
+            if block_count <= 0:
+                print(f"  Stage {idx}: no transformer blocks on {device}")
+                continue
+            start = running
+            end = start + block_count
+            running = end
+            print(f"  Stage {idx}: layers {start}-{end - 1} on {device}")
+            estimated_gb = device_block_bytes[idx] / (1024 ** 3)
+            print(f"           ≈{estimated_gb:.2f} GB of parameters")
+        print(
+            "  Final RMSNorm and LM head on "
+            f"{self.output_device} (stage {output_device_idx})"
+        )
+        output_gb = (device_block_bytes[output_device_idx] + norm_bytes + head_bytes) / (
+            1024 ** 3
+        )
+        print(f"           Estimated post-head load: ≈{output_gb:.2f} GB")
+
+    def _prepare_attention_mask(
+        self,
+        attention_mask: Optional[torch.Tensor],
+        batch_size: int,
+        seq_length: int,
+        device: torch.device,
+        dtype: torch.dtype,
+    ) -> Optional[torch.Tensor]:
+        """Prepare 4D attention mask from 2D mask (batch_size, seq_length).
+        
+        Returns a 4D mask suitable for scaled_dot_product_attention.
+        The mask should be additive (0 for attend, -inf for mask out).
+        """
+        if attention_mask is None:
+            return None
+        
+        # Convert 2D mask to 4D: (batch_size, seq_length) -> (batch_size, 1, seq_length, seq_length)
+        # Create causal mask
+        causal_mask = torch.triu(
+            torch.ones(seq_length, seq_length, dtype=torch.bool, device=device),
+            diagonal=1,
+        )
+        
+        # Expand attention_mask from (batch, seq) to (batch, 1, 1, seq)
+        expanded_mask = attention_mask[:, None, None, :].expand(batch_size, 1, seq_length, seq_length)
+        
+        # Combine: positions that are either causally masked OR padding should be masked
+        # attention_mask is 1 for attend, 0 for mask -> invert for additive mask
+        # Use a large negative value instead of -inf to avoid numerical issues in bfloat16
+        # -65504 is approximately the most negative value representable in float16
+        # Using a more conservative value for numerical stability
+        min_dtype = torch.finfo(dtype).min if dtype.is_floating_point else -1e9
+        mask_value = max(min_dtype, -65504.0)  # Clamp to avoid true -inf
+        
+        combined_mask = torch.where(
+            causal_mask | (expanded_mask == 0),
+            torch.tensor(mask_value, dtype=dtype, device=device),
+            torch.tensor(0.0, dtype=dtype, device=device),
+        )
+        
+        return combined_mask
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        **kwargs,  # Accept extra args from newer transformers (e.g., num_items_in_batch)
+    ) -> CausalLMOutput:
+        batch_size, seq_length = input_ids.shape
+
+        if self.pipeline_partitions:
+            first_device = self.pipeline_partitions[0][2]
+            hidden_states = self.embed_tokens(input_ids.to(first_device))
+            
+            # Prepare 4D attention mask
+            if attention_mask is not None:
+                attention_mask = self._prepare_attention_mask(
+                    attention_mask.to(first_device),
+                    batch_size,
+                    seq_length,
+                    first_device,
+                    hidden_states.dtype,
+                )
+            
+            cos, sin = self.rotary_emb(hidden_states, seq_length)
+
+            for start, end, device in self.pipeline_partitions:
+                if hidden_states.device != device:
+                    hidden_states = hidden_states.to(device)
+                rotary = (cos.to(device), sin.to(device))
+                attn_mask = attention_mask.to(device) if attention_mask is not None else None
+
+                for layer in self.blocks[start:end]:
+                    if self.gradient_checkpointing and self.training:
+                        hidden_states = torch.utils.checkpoint.checkpoint(
+                            layer,
+                            hidden_states,
+                            rotary,
+                            attn_mask,
+                            use_reentrant=False,
+                        )
+                    else:
+                        hidden_states = layer(hidden_states, rotary, attn_mask)
+
+            hidden_states = hidden_states.to(self.output_device)
+        else:
+            device = self.embed_tokens.weight.device
+            hidden_states = self.embed_tokens(input_ids.to(device))
+            
+            # Prepare 4D attention mask
+            if attention_mask is not None:
+                attention_mask = self._prepare_attention_mask(
+                    attention_mask.to(device),
+                    batch_size,
+                    seq_length,
+                    device,
+                    hidden_states.dtype,
+                )
+            
+            cos, sin = self.rotary_emb(hidden_states, seq_length)
+            rotary = (cos, sin)
+
+            for layer in self.blocks:
+                if self.gradient_checkpointing and self.training:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        layer,
+                        hidden_states,
+                        rotary,
+                        attention_mask,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = layer(hidden_states, rotary, attention_mask)
+
+        hidden_states = self.norm(hidden_states)
+        logits = self.lm_head(hidden_states)
+        
+        # Check for NaN in logits and handle gracefully
+        if torch.isnan(logits).any():
+            # Replace NaN with zeros to prevent cascading failures
+            logits = torch.nan_to_num(logits, nan=0.0, posinf=65504.0, neginf=-65504.0)
+
+        loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            if shift_labels.device != shift_logits.device:
+                shift_labels = shift_labels.to(shift_logits.device)
+            loss = F.cross_entropy(
+                shift_logits.view(-1, shift_logits.size(-1)),
+                shift_labels.view(-1),
+                ignore_index=-100,
+            )
+            # Handle NaN loss
+            if torch.isnan(loss):
+                loss = torch.tensor(0.0, device=loss.device, dtype=loss.dtype, requires_grad=True)
+
+        return CausalLMOutput(logits=logits, loss=loss)
+
+    @torch.no_grad()
+    def generate(
+        self,
+        input_ids: torch.Tensor,
+        max_length: int = 1024,
+        temperature: float = 1.0,
+        top_k: Optional[int] = None,
+        top_p: Optional[float] = None,
+        do_sample: bool = True,
+    ) -> torch.Tensor:
+        self.eval()
+        device = self.pipeline_partitions[0][2] if self.pipeline_partitions else self.embed_tokens.weight.device
+        input_ids = input_ids.to(device)
+        while input_ids.shape[1] < max_length:
+            chunk = input_ids[:, -self.config.max_position_embeddings :]
+            outputs = self.forward(chunk)
+            logits = outputs.logits[:, -1, :] / temperature
+
+            if do_sample:
+                if top_k is not None:
+                    top_values, _ = torch.topk(logits, min(top_k, logits.size(-1)))
+                    logits = logits.masked_fill(logits < top_values[:, [-1]], float("-inf"))
+                if top_p is not None:
+                    sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+                    cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+                    sorted_indices_to_remove = cumulative_probs > top_p
+                    sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+                    sorted_indices_to_remove[..., 0] = 0
+                    indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+                    logits = logits.masked_fill(indices_to_remove, float("-inf"))
+                probs = F.softmax(logits, dim=-1)
+                next_token = torch.multinomial(probs, num_samples=1)
+            else:
+                next_token = torch.argmax(logits, dim=-1, keepdim=True)
+
+            input_ids = torch.cat([input_ids, next_token.to(input_ids.device)], dim=-1)
+            if input_ids.shape[1] >= max_length:
+                break
+        return input_ids.to(device)
+
+
+AutoConfig.register("argonne2", ArgonneConfig)
+AutoModel.register(ArgonneConfig, ArgonneModel)
+AutoModelForCausalLM.register(ArgonneConfig, ArgonneModel)
+
+# Backwards compatibility exports
+CausalSelfAttention = GroupedQueryAttention
+MLP = SwiGLUMLP

model.safetensors.index.json

@@ -0,0 +1,226 @@
+{
+  "metadata": {
+    "total_size": 22147464000
+  },
+  "weight_map": {
+    "embed_tokens.weight": "model-00001-of-00005.safetensors",
+    "lm_head.weight": "model-00001-of-00005.safetensors",
+    "blocks.0.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
+    "blocks.0.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
+    "blocks.0.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.1.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.1.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.1.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.2.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.2.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.2.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.3.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.3.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.3.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.4.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.4.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.4.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.5.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.5.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.5.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.6.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.6.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.6.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.7.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.7.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.7.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.8.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.8.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.8.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.9.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.9.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.9.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.10.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
+    "blocks.10.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.10.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.11.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.11.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.11.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.12.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.12.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.12.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.13.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.13.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.13.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.14.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.14.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.14.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.15.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.15.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.15.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.16.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.16.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.16.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.17.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.17.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.17.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.18.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.18.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.18.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.19.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.19.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.19.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
+    "blocks.20.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.20.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.20.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.21.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.21.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.21.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.22.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.22.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.22.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.23.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.23.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.23.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.0.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.0.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.1.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.1.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.2.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.2.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.3.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.3.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.4.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.4.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.5.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.5.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.6.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.6.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.7.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.7.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.8.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.8.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.9.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.9.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.10.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.10.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.11.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.11.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.12.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.12.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.13.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.13.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.14.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.14.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.15.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.15.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.16.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.16.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.17.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.17.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.18.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.18.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.19.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.19.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.20.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.20.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.21.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.21.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.22.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.22.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.23.attn.q_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.23.attn.o_proj.weight": "model-00004-of-00005.safetensors",
+    "blocks.0.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.0.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.1.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.1.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.2.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.2.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.3.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.3.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.4.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.4.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.5.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.5.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.6.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.6.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.7.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.7.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.8.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.8.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.9.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.9.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.10.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.10.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.11.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.11.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.12.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.12.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.13.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.13.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.14.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.14.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.15.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.15.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.16.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.16.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.17.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.17.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.18.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.18.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.19.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.19.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.20.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.20.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.21.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.21.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.22.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.22.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.23.attn.k_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.23.attn.v_proj.weight": "model-00005-of-00005.safetensors",
+    "blocks.0.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.0.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.1.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.1.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.2.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.2.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.3.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.3.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.4.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.4.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.5.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.5.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.6.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.6.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.7.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.7.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.8.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.8.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.9.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.9.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.10.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.10.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.11.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.11.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.12.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.12.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.13.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.13.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.14.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.14.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.15.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.15.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.16.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.16.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.17.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.17.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.18.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.18.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.19.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.19.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.20.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.20.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.21.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.21.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.22.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.22.post_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.23.input_norm.weight": "model-00005-of-00005.safetensors",
+    "blocks.23.post_norm.weight": "model-00005-of-00005.safetensors",
+    "norm.weight": "model-00005-of-00005.safetensors"
+  }
+}

special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,207 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- messages[0]['content'] }}\n    {%- else %}\n        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}\n    {%- endif %}\n    {{- \"\\n\\n# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n    {%- else %}\n        {{- '<|im_start|>system\\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {{- '<|im_start|>' + message.role }}\n        {%- if message.content %}\n            {{- '\\n' + message.content }}\n        {%- endif %}\n        {%- for tool_call in message.tool_calls %}\n            {%- if tool_call.function is defined %}\n                {%- set tool_call = tool_call.function %}\n            {%- endif %}\n            {{- '\\n<tool_call>\\n{\"name\": \"' }}\n            {{- tool_call.name }}\n            {{- '\", \"arguments\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- '}\\n</tool_call>' }}\n        {%- endfor %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "model_max_length": 1000000000,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}